Deep learning solution for in-line microscopy monitoring and particle size analysis

We introduce a novel deep learning based image analysis method for in- line microscopy monitoring and particle size analysis. This methodology allows us to quantify the size of crystals from low-quality Particle Vision and Measurement (PVM) images taken during the crystallization process.PVM technology captures particles and particle mechanisms as they exist in the process. Due to the variability of concentrations and the size of the crystals, the appearance of the PVM images is hazy and the individual crystals are difficult to recognize, which makes the analysis of particle size challenging.The proposed method consists of two major components: a novel way to generate training image and label pairs with a high level of credibility via a Cycle-Consistent Adversarial Network (CycleGAN), and a Mask-RCNN model trained with the generated data for the particle detection task. Compared with manual labeling and conventional supervised deep learning approaches, which are normally labor-intensive and error-prone, the method is more automatic and objective since it can benefit from highly realistic training data synthesis.The method has been applied to images of two particle morphologies with varying particle sizes and solids concentrations (typical concentrations of actual industrial conditions). The results demonstrate that the method can successfully perform relative size comparisons for the lower end of the solids concentrations, suggesting that this method could be used to monitor crystallization mechanisms such as growth, agglomeration, and breakage in the lower end of the industrial concentration spectrum.